Platform · How it works
Everything between hello and done.
Every agent answers fast, survives a provider outage, keeps to your rules and can run on your own model accounts. Here is the runtime, opened up.
The runtime
- Solid line: Voice
- Dashed line: Written
- Numbered marks: Explained below
Channels
- Phone, voice
- Web, voice and written
- Email, written
- SMS, written
- Messaging apps, written
Instead of the three stages below
OrRealtime speech-to-speechOne model hears, thinks and speaks
Listening, carries Voice
Speech-to-textfirst choice: Deepgram
+1 backup, in rank order: ElevenLabs
Turn detector1
End of turn
Thinking, carries Voice and written
Language modelfirst choice: OpenAI2
+2 backups, in rank order: Anthropic, Google
Tools3
100+ connectors and integrations
Speaking, carries Voice
Text-to-speechfirst choice: ElevenLabs
+2 backups, in rank order: Cartesia, Fish Audio
Pronunciation
VIN → “V-I-N”
Reply
On the line
In writing
Storage
Google Cloud · us-east4
Northern Virginia, United States
Key vault4
Google Secret Manager
Keys sealed with AES-256-GCM
- 1LatencyAnswers the moment they’re done.: Latency and turn-taking
- Pause: Keeps listening
- Turn over: Answers
- Caller cuts in: Gives way
- 2ReliabilityThe call rides out a failed provider.: Reliability
- OpenAI: Timeout
- Anthropic: Took the turn
- Google: In line
- 3SecurityNo caller talks it past your rules.: Security
- Find an open slot: Allow
- Refund the deposit: Ask
- Delete a customer record: Block
- 4Model freedomYour own accounts, at every stage.: Model freedom (BYOK)
- Deepgram: active, key ending 31b0
- OpenAI: active, key ending 3kQ7
- ElevenLabs: active, key ending f7K2
Building on it? The API, the MCP server and webhooks
Watch Zumu answer one of your calls.
In half an hour, the people who built it take that call through every part above, to the booking it ends in.
Listening · Thinking · Speaking
